

Changelog
=========

Version 4.0.1
-------------

build/install
~~~~~~~~~~~~~

    **features**

    - Add Cygwin support.

    - Update config.guess and config.sub to the latest versions.

    **build fixes**

    - Fix pkg-config files for custom paths.

    - Set LD/LDXX unconditionally to CC/CXX, since LDFLAGS/LDXXFLAGS from
      ./configure rely on it.

    - macOS: use libdir instead of RPATH for the install_name.


Version 4.0.0
-------------

general
~~~~~~~

    **sync soversion and major_version**

    - The added number formatting feature requires an ABI change, hence the
      increase to SOVERSION=4.

    - Packagers outside of the Linux distributions sometimes use the major
      version number as the equivalent of SOVERSION on their platforms and
      have an incorrect SOVERSION for 2.5.1, which requires SOVERSION=3.

    - While SOVERSION is not required to match the major version number
      (example: glibc), mpdecimal will from now on take the path of least
      resistance and always use SOVERSION=MPD_MAJOR_VERSION.

    - The jump to 4.0.0 should also remind users that a C++ library is
      available.

build/install
~~~~~~~~~~~~~

    **features**

    - Support for out-of-tree build.

    - Support for pkg-config.

    - Unix: support for Loongson.

    - Unix: support for CheriBSD.

    - Compilers: support for icx, icpx, ibm-clang_r, ibm-clang++_r, CompCert,
      clang-cl and emscripten.

    - Windows: support for MSYS2/MinGW.

    - MSVC: the build now uses /O2 /DNDEBUG.

    - MSVC: new arm64/arm32 cross build scripts.

    - AIX: the shared libraries are now installed as versioned objects, e.g.,
      shr4.o, shr4_64.o.

    - New ``./configure`` switches:

      ``--enable-static``: enable/disable the build of the static libraries
      (default: enabled).

      ``--enable-pc``: enable/disable the install of the pkgconfig files
      (default: enabled).

      ``--enable-doc``: enable/disable the install of the documentation
      (default: enabled).

    - New ``./configure`` behavior:

      On multilib platforms like AIX that default to 32-bit the 64-bit build can
      be forced with a single ``./configure MACHINE=uint128`` (ibm-clang, gcc)
      or ``./configure MACHINE=ansi64`` (xlc) rather than explicitly setting
      ``CFLAGS``, ``LDFLAGS``, ``CXXFLAGS`` and ``LDXXFLAGS``.

      Note that on most other (multilib) platforms the default is 64-bit and the
      optimal configuration is chosen by just using ``./configure``.

    - New man pages direct users to the mpdecimal-doc package or the online
      HTML documentation.

    **tests**

    - The C++ tests now automatically skip a small number of bignum tests if
      the std::thread stack size (which cannot be altered) is less than 512K.

      For all other tests a thread stack size of roughly 50K is sufficient.

    - Files in the tests/testdata_dist directory have been significantly
      abbreviated in order to facilitate testing on slower platforms. The
      tests take 5s on a modern x86_64 platform.

      The previous tests have been moved to the separate mpdecimal-testdata
      package (replace tests/testdata_dist with the renamed mpdecimal-testdata
      directory).

    **toolchain issues**

    - AIX: The preferred C/C++ compilers are ibm-clang_r and ibm-clang_r++.
      g++ has sporadic C++11 threading issues that only show up in long running
      programs (or short ones with a sufficient number of repetitions):

         https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98390

      If you only use libmpdec (``--disable-cxx``), gcc and xlc are very stable.

    **changed**

    - The ``--enable-profile`` configure option has been removed due to fragile
      integration with a sequence like ``make && make install``. The new method
      is ``make profile && make install``.

    **removed**

    - The prebuilt HTML documentation is now in the separate mpdecimal-doc
      package, which gives distributions that reject prebuilt documentation
      the option to disregard it and use the new man pages.

libmpdec
~~~~~~~~

    **features**

    - Add the "z" format specifier (coerce negative zeros to positive).

    - In extremely rare cases the transcendental functions (exp, ln, log10) did
      not set the Subnormal/Underflow flags. The reason is that in the case of
      an exponent boundary the Ziv correction loop for correct rounding requires
      very few iterations to arrive at the correctly rounded result, but may
      need many more iterations to arrive at the correct flags.

      In these cases, Subnormal/Underflow is not very informative, so the status
      quo was to skip the extra iterations.

      Version 4.0.0 now specializes exponent boundary cases and uses up to five
      additional iterations to set Subnormal/Underflow. The refactored code has
      no speed penalty on average;  in fact, in the deccheck tests (random tests
      with a bias towards corner cases) it is slightly faster.

      No cases have been found where more than two additional iterations are
      required, but they may exist.

    **reliability fixes**

    - mpd_qset_string_exact(), mpd_qset_i64_exact() and mpd_qset_u64_exact()
      can now be called with a nonzero status.  Previously, the functions
      could return NaN/Invalid_operation in that case.

      This is listed under "reliability fixes" since there is no possible
      scenario under which these functions would legitimately be called with
      a nonzero status.

libmpdec++
~~~~~~~~~~

    **features**

    - Add input validation for Decimal.shiftl(), Decimal.shiftr() and
      Decimal::ln10().


Version 2.5.1
-------------

libmpdec
~~~~~~~~

    **features**

    - New functions for conversion between mpd_t and a decimal triple
      with a split uint128_t coefficient.

    - All symbols in the header files now have an MPD_* or mpd_* prefix.

    - libmpdec/.objs/symbols64.exp and libmpdec/.objs/symbols32.exp contain
      all symbols of the exported API.  The files are almost identical, but
      mpd_qsset_i64, mpd_sset_i64, mpd_qsset_u64 and mpd_sset_u64 are only
      available in the 64-bit build.

      libmpdec/.objs/libmpdec.imp is the unified 64/32-bit import file
      for AIX.

    - The new test target *make check_local* does not attempt to download
      dectest.zip.  This is useful for packagers whose infrastructure does
      not allow downloading during the testing phase.

      *make check* is still preferable because it runs all tests.

    **style fixes**

    - Apply several clang-tidy suggestions (high diagnostic level).

    **build features**

    - AIX: full support for xlc/gcc builds in ./configure.
    - full support for AIX-style shared libraries.
    - full support for AIX multilib header.

    **build fixes**

    - Update config.guess and config.sub in order to support IBM Power
      architectures.

    - Fix false positive _FORTIFY_SOURCE warning on Fedora.

libmpdec++
~~~~~~~~~~

    **features**

    - New methods for conversion between Decimal and a decimal triple
      with a split uint128_t coefficient.

    - decimal::Context now uses default constructors wherever possible.

    - AIX: The tests now have a --pthread option. The default thread stack
      size on AIX (96K) makes std::thread unusable for runtest.cc, which needs
      around 300K.

    - Windows: The DLL can now be built. Since DLLs do not support the
      C++11 extern thread_local context, each translation unit gets its
      own static reference to the thread_local context.  This should
      preserve the semantics at the expense of clarity.

      Note that this scheme is specific to the DLL build; the static
      library build on Windows as well as all other operating systems
      (except for OpenBSD and Solaris) use the C++11 extern thread_local
      context.

    - OpenBSD, Solaris: The shared library now works using the above scheme.
      For simplicity, both the shared and static libraries use the workaround.

    **style fixes**

    - Apply several clang-tidy suggestions (high diagnostic level).

    **bug fixes**

    - DecimalException now inherits from std::runtime_error instead
      of std::exception.  This automatically provides a nothrow copy
      constructor.

    - Use mpd_free() instead of free(). This only affects applications using
      the libmpdec custom allocators.

    **build fixes**

    - Undefine some additional math.h macros for non-compliant C++11
      implementations.

    - libmpdec++ is now linked against libmpdec.


Version 2.5.0
-------------

    **New: libmpdec++**

       libmpdec++ is a new C++ library around libmpdec. It frees users from
       manual memory management and allows cleaner code using inline operators.

       libmpdec++ has a thread local context, so inline operators work
       seamlessly with threaded code.

    **features**

    1) New functions *mpd_qset_string_exact*, *mpd_qset_i64_exact* and
       *mpd_qset_u64_exact*.

    2) For very large precisions like *MPD_MAX_PREC* libmpdec failed with
       *MPD_Malloc_error* even when the result was exact and required far
       fewer digits.

       Now libmpdec retries the operation with the smallest possible
       precision for exact results.

    3) ./configure now has the *--enable-profile* option for easier profile
       guided optimization builds.  Using PGO, recent compilers produce
       better optimized libraries with speedups in the order of 20%.

    4) C++ use: All headers now assume at least C++11, so a couple of hacks
       have been removed from the headers.  C++ compatibility is fully
       tested in libmpdec++.

    **behavior changes**

    1) The functions *mpd_ceil*, *mpd_floor* and *mpd_trunc* now set
       *MPD_Invalid_operation* when the input is *NaN* or *Infinity*.

       Previously they returned special values unchanged. That behavior was
       inspired by the similar *to_integral* function from the specification,
       but does not seem useful for the above functions.

    **build fixes**

    1) OS X: Linking has been completely reworked and now uses .dylib.

    2) Windows: vcstdint.h has been removed, Windows now supports stdint.h.

    **bug fixes**

    1) A couple of quiet functions have been made resilient against being
       called with a "dirty" status. Note that it is not recommended in
       general to call quiet functions with any status other than 0.

    2) In runtest.c, fma_eq.decTest and powmod_eq.decTest operands were
       accidentally reversed for the *op_eq_eq* tests.

       Since both the code and the tests assumed reversed operands, the tests
       were correct.  However, for readability this situation has been fixed.


Version 2.4.2
-------------

    **build fixes**

    1) ICC/Windows: the optimized x86 build requires -fp-model=precise.

    2) OS X: the linker requires -dynamiclib and -install_name for building
       a shared library.


Version 2.4.1
-------------

    **build fixes**

    1) The __uint128_t detection in ./configure has been fixed .  Failure to
       detect the option resulted in building the significantly slower ANSI
       target on non-x86/amd64 platforms.

    2) Use -fPIC instead of -fpic to fix a build failure on SPARC platforms.

    3) Split the tests into a faster "make check" and a slower "make check_alloc".
       The latter tests allocation failures but is too slow on older machines.

    4) Generate detailed test output for better feedback on slower machines.

    5) The static library is now built without -fPIC, which is significantly
       faster at least on x86.  Both the static and the shared library are
       now tested separately.

    **bug fixes**

    1) PEP 3101 formatting:  With the '%' format type, a trailing percent sign
       is now also added for infinities and NaNs.


Version 2.4.0
-------------

    **features**

    1) Faster integer to string conversion.

    2) mpd_qln(), mpd_qlog10() and mpd_pow() are now thread-safe.

    3) All functions that take or return C integers are now available in
       both the 64-bit and the 32-bit builds.

    4) Support for cross-compiling.

    5) Scripts for Visual Studio builds.

    **code improvements**

    1) This version is exactly the same as the version shipped with Python-3.3+.
       Large portions of the code have been refactored in order to facilitate
       proofs.  Many ACL2 proofs have been added.

    **removed**

    1) The Python module has been removed from mpdecimal, since both libmpdec
       and cdecimal are included in Python-3.3+.

    2) The large test suite against decNumber as well as the multi-precision
       tests against gmp have been removed, but will be available in a separate
       package.  Naturally these tests are still run as part of the release
       process.


Version 2.3
-----------

    **features**

    1) New test suite with comprehensive tests against decNumber.

    2) Full support for compilers without uint64_t (tested with CompCert).

    **bug fixes**

    1) If ROUND_FLOOR is set and the operand is zero, the functions
       mpd_plus() and mpd_minus() have special cases for the sign of
       the result.


Version 2.2
-----------

    **build process**

    1) configure: append CFLAGS to CONFIG flags.

    2) Makefile: use includedir, libdir, datarootdir, datadir, docdir, DESTDIR.

    **workarounds for toolchain bugs**

    1) Enable workaround for a gcc miscompilation. See:

           `http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46491 <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46491>`_

    2) Enable workaround for the glibc _FORTIFY_SOURCE/memmove bug, which is
       exposed by gcc-4.6. See:

           `http://sourceware.org/ml/libc-alpha/2010-12/msg00009.html <http://sourceware.org/ml/libc-alpha/2010-12/msg00009.html>`_

    **features**

    1) Make PPRO inline assembly PIC-compliant (for the dynamic library).


Version 2.1
-----------

Version 2.1 was never released, but escaped into the wild via the Makefile
and setup.py in cdecimal-1.97-rc2.tar.gz. Both files already had that version
number.

    **features**

    1) Code coverage increased to 100%. This includes every possible
       allocation failure.

    2) Switch build process to ./configure.

    3) Makefile targets for creating coverage reports.

    **bug fixes**

    1) mpd_qget_uint, mpd_qget_u64, mpd_qget_u32 did not raise for
       negative input.

    2) Handle allocation failures in _mpd_fntmul under extreme conditions.



Version 1.2.1
-------------

    **bug fixes**

    1) With MACHINE=ansi64, the macros BSR and BSF used x86 assembly.
       This caused compilation to fail on non-x64 platforms.


Version 1.2
-----------

    **features**

    1) Support for compilers with __uint128_t
       (option MACHINE=ansi64).

    2) Support for other 64-bit compilers
       (option MACHINE=ansi64c32).

    3) Support for legacy compilers without uint64_t
       (option MACHINE=ansi-legacy).

    4) Slightly different build process (please read INSTALL.txt).

    5) If clamp=1, the maximum payload length of a NaN is prec-1.

    **bug fixes**

    1) Fix for mpd_qround_to_int, which did not handle digits
       exceeding the context precision correctly in all cases.

    2) In rare corner cases Underflow was not set in
       transcendental functions.



